Machine Learning Framework for Alzheimer’s Disease Protein Classification using Sequence-Derived and Network Topological Features

Alzheimer's disease (AD) is a neurodegenerative disorder whose molecular complexity cannot be adequately captured using sequence-level features alone. This paper proposes AlzGenPred+, a biologically grounded and interpretable computational framework for classifying AD-associated protein sequences. The proposed model integrates amino acid sequence representations with protein–protein interaction (PPI) network topology, processing 1,000 proteins of fixed length (200 amino acids) derived from a proteomic dataset. Protein sequences are encoded using overlapping 3-mer representations, followed by dimensionality reduction via Principal Component Analysis (PCA) to 32 principal components. A Multi-Layer Perceptron (MLP) learns latent embeddings supervised by PPI network centrality measures (degree, betweenness, closeness centrality, and clustering coefficient). Final binary classification (AD-associated vs. non-associated) is performed by a Gradient Boosting Classifier, combining network features and deep embeddings in a 36-dimensional fused vector. Evaluation demonstrates discriminative performance with AUC = 0.546 (network-only), F1-Score = 0.85, and Pearson correlation r = 0.81 in embedding correlation analysis. A web-based predictive platform enables real-time probabilistic evaluation of novel protein sequences. The proposed framework advances translational bioinformatics by offering a scalable, interpretable, and biologically meaningful system for AD-associated protein discovery.